Mm/pure gradient descent

[mmcleod89]

• Add pure gradient descent method
• takes a function f(x,y) [likelihood] and g(x) [prior]
• Uses FISTA update but this should probably be a flag
• Need to settle on an appropriate interface
• Convergence fn should be passed into setup, possibly with a default. This seems to be the main point of difference between Algorithm and ImagingAlgorithm classes. (I think the imaging algorithm classes might be redundant or could inherit from the algorithm classes to avoid all the unnecessary forwarding of arguments.)
• Needs to be tested with real images & CNN model